【作者】 吴文钦；

【导师】 杨斌；

【作者基本信息】 西南交通大学 ， 计算机应用， 2003， 硕士

【摘要】 近年来车辆牌照自动识别技术在智能交通领域中的运用非常的广泛。在已经实现的系统中还存在着识别性能不稳定，对垂直倾斜角度过大情况不能纠正，以及针对于字符识别而设计的分类器识别率不高并且训练过程过长等缺点，需要科研人员继续对系统进行完善。 一个完整的车辆牌照识别系统工作流程一般由三部分组成：图象输入，牌照定位与字符识别。本文主要讨论软件部分的工作，包括牌照定位和字符识别。围绕以上任务，本文对牌照倾斜角度的计算方法、对垂直倾斜角度过大情况下对牌照图象的纠正方法，车牌精确定位方法、字符识别的方法和多分类器的集成方法等作了深入的研究和探讨。主要研究工作包括： 在对牌照倾斜角度的计算方法中率先提出使用Canny算子作边缘检测。通过实验证明Canny算子比其它检测算子如Sobel等对车牌边框的检测效果要好。 提出了在对垂直倾斜角度过大的情况下使用简单空间变换的方法来纠正图象的畸变而对垂直倾斜角度不大的情况则不作为的新策略。这种分别对待的目的是使字符尽量不出现锯齿形，同时可识别度增高。 在车牌精确定位环节中阐述了两种通用的方法：“基于角检测的牌照字符区域定位”，和“基于边缘检测和快速Hough变换的区域定位”。而作者提出一种全新方法—基于几何学的算法，它根据几何坐标变换的转换关系和对车牌中心坐标的估计计算，以非常精确的公式推导出精确提取牌照区域的四个顶角坐标。与前两种方法比较，后一种方法计算速度快，抗倾斜能力强。 在字符识别环节中阐述了结构识别方法，统计识别方法，结构与统计相结合的组合识别方法，以及神经网络分类器识别的方法。对HMM(隐马尔可夫模型)方法的应用现状作了简单介绍。通过对字母数字识别的分类器模块设计方案的实验，指出采用置信度的组合多分类器集成的方法能够有效提高识别率。 对机器学习领域当前的最新研究热点统计学习理论和支持向量机(SVM)方法在字符识别上的应用作了研究与探讨。支持向量机在模式识别中的对有限样本的要求，以及不同的核函数可以得到类似的识别效果，支持向量机的训练时间周期较短，识别率较高等方面都表现出比其它机器学习方法更优异的性能。 系统在实现过程中一方面选择采用了来自各种学术文献上的较为优秀的成熟方法，另一方面在某些环节采用了性能更好的新算法，并对引入某些新理论加以应用作了尝试。系统的稳定性也在开发过程中始终被强调。 西南交通大学硕士研究生学位论文 第日页 将来进一步的工作包括以下几个方面：将基于SVM的字符分类器引入系统中，提高系统的整体识别率。增加图象预处理功能，对光照过强或过弱条件下所拍摄图象进行质量提升。更多还原

【Abstract】 The techniques of Car License Plate Recognition is used widely in the ITS field recently.There are some problems still existed in the current working system:the performance of system is not stable enough; the remedy method of distorted image is not adopted in the case of vertical incline grade is large; the recognition rate of character classifier is not high enough to meet the need and the training period of character classifier is too long . All of these need the researchers to better the algorithms and optimize the system.An whole Car License Plate Recognition System consists of three components: image inputting^icense retrieving^character recognition. This thesis focuses on the software part of the system including "license retrieving" , " character recognition" .In order to solve the problems described above, some deep researches including the algorithm of the computing of incline grade of license plate, the remedy method in the case of vertical incline grade is large, the exact plate retrieving algorithm, the character recognition method, and the combination method of multi-classifiers, have been carried out in this thesis.Main research work of this thesis includes:Canny arithmetic operator of edge detection is introduced firstly in the algorithm of the computing of incline grade of license plate. The experiment result proves the performance of Canny operator is more excellent than the traditional operator such as Sobel.A "simple space-transformation" method is proposed to remedy the image distortion in the case of vertical incline grade is large whereas nothing is done in the case of vertical incline grade is small in this thesis. This purpose of novel method of discriminated treatment is to improve the recognition rate, and reduce the zigzag occurrence of characters.Two universal methods of exact plate retrieving are presented. One is exact plate retrieving method based on corner detection. The otheris exact plate retrieving method based on line edge inspecting. And a new method is proposed here based on the geometry. According to the transformation rules of the geometry coordinates before and after the image rotation and the estimated computation of the coordination of the plate center, a formula could be deduced , and the four corners are marked. Compared with the two universal methods , the geometry-based method has a shorter computing-time consumption, and a stronger ability of anti-incline of plate.In the stage of character recognition , a lot of methods are presented in this thesis, including the structure recognition method , the statistical recognition method, the structure and statistical combination recognition method, and artificial neural network classifier recognition method. Hidden Markov Module and its application in character recognition are presented in brief. It is proved here that the method of combination of multi-classifiers with the confidence computing could improve the recognition rate via an alphabet and digits recognition classifier module scheme experiment.Some researches and discussion have been carried out on the application of Statistical Learning Theory and Support Vector Machine in character recognition. SLT and SVM are the new hotspots in machine learning field. SVM represents better performance than most other machine learning method because of the requirement for limited number training samples, different kernel function could deduce the approximate recognition result, shorter training-period interval, higher recognition rate .On the one hand the "SeeCar" system have adopted many known excellent methods cited from the domestic and international published paper, on the other hand some new algorithm with better performance have been put up . The stability of system has been emphasized during the whole developing period .The next step work will be carried out in the future including the following aspects: A SVM-based character classifier will be introduced in the recognition module with the aim to ehance the whole recognitionrate; A pre-更多还原